Trenton Bricken

Trenton Bricken is an Anthropic researcher who studies what happens inside language models, first through interpretability and now through alignment auditing. A 2020 Duke graduate with a self-made major in biological and artificial intelligence, Bricken wrote a NeurIPS 2021 paper with Cengiz Pehlevan that relates Transformer attention to Sparse Distributed Memory. At Anthropic, the first-listed authorship of Towards Monosemanticity in 2023 brought a sparse-autoencoder method for extracting interpretable features from a one-layer transformer, and the 2024 Scaling Monosemanticity paper applied the approach to Claude 3 Sonnet. In March 2025, they defended a Harvard PhD in systems biology and co-authored Auditing Language Models for Hidden Objectives. They are a Member of Technical Staff on Anthropic's Alignment Science team, where the work is enabling Claude to automatically audit and detect misalignment.

Now at: Anthropic
Badges: Bridge

Timeline

2020

A self-made major at Duke

In May 2020, they graduate from Duke University as a Robertson Scholar with a self-made major in Minds and Machines: Biological and Artificial Intelligence. At Duke, they spend a year in Michael Lynch's lab using machine learning to design CRISPR guide RNAs, and afterwards they affiliate with Debora Marks's lab at Harvard Medical School to apply deep learning to protein design.
2021

Attention and sparse memory

In 2021, they publish Attention Approximates Sparse Distributed Memory with Cengiz Pehlevan at NeurIPS. The paper shows that Transformer attention can be closely related, under certain data conditions, to Kanerva's Sparse Distributed Memory, and it confirms those conditions in pre-trained GPT2 models.
By 2023

A resident at Anthropic

By 2023, they are at Anthropic, where they begin as Tristan Hume's resident after the two meet over a paper on sparsity in networks, and then convert to full-time. They also work as a visiting researcher at Berkeley's Redwood Center for Theoretical Neuroscience, collaborating with Bruno Olshausen on vector symbolic architectures and sparse coding.
October 2023

Towards Monosemanticity

In October 2023, they are a first-listed author and core contributor of Towards Monosemanticity: Decomposing Language Models With Dictionary Learning, an Anthropic paper. Using a sparse autoencoder, the authors extract a large number of interpretable features from a one-layer transformer.
2024

Scaling Monosemanticity

In March 2024, they appear on the Dwarkesh Podcast with Sholto Douglas, where host Dwarkesh Patel describes them as an Anthropic researcher working on mechanistic interpretability. In May 2024, they are among the authors of Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet.
March 2025

Auditing for hidden objectives

In March 2025, they are among the co-authors of Auditing Language Models for Hidden Objectives. The team trains a language model with a hidden objective, and in a blind auditing game three of four teams uncover it.
March 2025

A Harvard systems biology PhD

In March 2025, they give their PhD defense, completing a Harvard PhD in Systems Biology. The thesis, Sparse Representations in Biological and Artificial Neural Networks, comes from the Kreiman Lab with support from an NSF Graduate Research Fellowship.
As of October 2026

Alignment auditing at Anthropic

In July 2025, they are first-listed author of Building and evaluating alignment auditing agents, which develops three agents that autonomously perform alignment auditing tasks. In February 2026, they post Cross-Architecture Model Diffing with Crosscoders on arXiv with Thomas Jiralerspong. They are a Member of Technical Staff on Anthropic's Alignment Science team, working on enabling Claude to automatically audit and detect misalignment.